MGR
★

India,
2014-05-22 07:45

Posting: # 12981
Views: 8,005

## SAS code for 2-stage design [Two-Stage / GS Designs]

Hi All,

We are going for a 2stage approach for EMA using Potvin B method.
Is this type of analysis can be done by WinNonlin 5.2 or 6.3? Moreover is there any articles/guidelines for the SAS code?

Regards,
MGR
d_labes
★★★

Berlin, Germany,
2014-05-22 08:17
(edited by d_labes on 2014-05-22 08:33)

@ MGR
Posting: # 12982
Views: 7,388

## SAS code for 2-stage design

Dear MGR,

» We are going for a 2stage approach for EMA using Potvin B method.

Welcome to the club .

» Is this type of analysis can be done by WinNonlin 5.2 or 6.3?

You may find some material in his lectures also, some decided to the 2-stage evaluation.

» Moreover is there any articles/guidelines for the SAS code?

AFAIK there is no decided article/guideline containing SAS code in this respect. But it should not that hard to code it:

For stage 1 use the code for classical ABE for a 2x2 cross-over, but instead of calculating 90% CIs derive the CIs to the confidence level (1-2*0.0294) = 94.12%.
The power inspection step and the sample size estimation are a little bit hairy in SAS. Don't really know if Proc Power can be used for that. I suggest to use some tool external to SAS. You may guess what I recommend (if you ever have seen or heard about): R-package PowerTOST.

For stage 2 evaluation (if a second stage becomes necessary) use a model with the effects:
• treatment
• stage
• period within stage
• sequence
• stage*sequence interaction (according to EMA Q&A)
• subject nested within sequence and stage
and derive again 94.12% CIs.

Here my SAS code for that:
Proc GLM data=yourdata;   class treatment period stage sequence subject;   model &lnval = treatment stage period(stage) sequence                  stage*sequence /*acc. to EMA Q&A*/                  subject(sequence*stage) / CLparm alpha=0.0588 ss3;   ODS output estimates=_est2;   estimate 'T-R' treatment -1 1;   where stage in(1 2); run; quit;
That is it. Not so difficult, isn't it?
Hope this helps.

BTW: There is a new revision of the EMA Q&A quite recently (09-May-2014) out there:
EMA/618604/2008 Rev. 9

Regards,

Detlew
Helmut
★★★

Vienna, Austria,
2014-05-22 12:24

@ MGR
Posting: # 12986
Views: 8,343

## Phoenix/WinNonlin code

Hi MGR,

» Is this type of analysis can be done by WinNonlin 5.2 or 6.3?

WinNonlin 5.2 is history (released in 2007, end of support of v5.2.1 was already in June 2011). In the following the procedure in Phoenix/WinNonlin’s 6.3 BE tool (EMA’s all fixed effects method and the addi­tional term in the second stage according to the Q&A-document Rev.7) and Example 2 from Potvin’s paper:
• Stage 1
Map variables as usual.
Navigate to [Options] and change the Confidence Level to 94.12%.
Navigate to [Random 1] and delete Subject(Sequence).
Navigate to [Fixed Effects], modify the model specification to Sequence+Treatment+Period+Subject(Sequence).
You should get:
Ratio_%Ref:    108.76 CI_User_Lower:  92.93 CI_User_Upper: 127.28
With this coding we don’t get the intrasubject-CV in the output (I think it will be available in the next release this summer).
Output DataFinal Variance ParametersSend ToDataData Wizard:
Add TransformationCustomNew Column Name: CVintraFormula: sqrt(exp(Estimate)-1)
You should get:
CVintra: 0.1821 (=18.21%)

• Stage 2
Map variables as usual; additionally map Stage to Classification.
Navigate to [Options] and change the Confidence Level to 94.12%.
Navigate to [Random 1] and delete Subject(Sequence).
Navigate to [Fixed Effects], modify the model specification to Stage+Sequence+Sequence*Stage+Sequence*Stage*Subject+Period(Stage)+Treatment.
You should get:
Ratio_%Ref:    101.45 CI_User_Lower:  88.45 CI_User_Upper: 116.38
These results agree with the ones reported by Potvin et al.

Intermediate power after stage 1 and sample size estimation for stage 2 in R with PowerTOST:
• library(PowerTOST) power.TOST(alpha=0.0294, theta0=0.95, CV=0.1821, n=12, design='2x2')
You should get:
[1] 0.5253565
Since 52.54% < 80%, initiate the second stage (if ≥80% stop; study failed to show BE).

• sampleN.TOST(alpha=0.0294, targetpower=0.80, theta0=0.95, CV=0.1821, design='2x2')
You should get:
+++++++++++ Equivalence test - TOST +++++++++++             Sample size estimation ----------------------------------------------- Study design:  2x2 crossover log-transformed data (multiplicative model) alpha = 0.0294, target power = 0.8 BE margins        = 0.8 ... 1.25 Null (true) ratio = 0.95,  CV = 0.1821 Sample size (total)  n     power 20   0.829292
Perform the second stage in 20–12=8 subjects like in the example.
BTW, recently it was shown* that EMA’s modification of the model is irrelevant for the BE assess­ment – which is not sur­prising, since Sequence*Stage is a between-subject term. If anybody has a clue why EMA introduced it, please enlighten me. Furthermore, there is not difference in results irrespective whether subjects are treated as a fixed or random effect (unless T/R >1.20).

• Karalis V, Macheras P. On the Statistical Model of the Two-Stage Designs in Bioequivalence Assessment. J Pharm Pharmacol. 2014;66(1):48–52. doi 10.1111/jphp.12164

Cheers,
Helmut Schütz

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
yjlee168
★★

Kaohsiung, Taiwan,
2014-05-23 21:17
(edited by yjlee168 on 2014-05-24 08:35)

@ Helmut
Posting: # 13001
Views: 7,076

## maybe R too?

Dear Detlew & Helmut,

Sorry to cut in. Thank you both for the great summary of the two-stage approach. It sounds like that it is possible to implement all these using R or into bear together with PowerTOST.

» [...]
» You should get:
» CVintra: 0.1821 (=18.21%)

OK, the example is from WLN. CVintra = 18.21% (for Cmax?) with 12 subjects. So it is not a HVD, is it? Does EMA or FDA accept the two-stage BE if it is the case?

» [...]
» sampleN.TOST(alpha=0.0294, targetpower=0.80, theta0=0.95, CV=0.1821, design='2x2')

Nice explanations and thank you both again.

All the best,
---Yung-jin Lee
bear v2.8.4:- created by Hsin-ya Lee & Yung-jin Lee
Kaohsiung, Taiwan http://pkpd.kmu.edu.tw/bear
Helmut
★★★

Vienna, Austria,
2014-05-23 22:39

@ yjlee168
Posting: # 13002
Views: 7,159

## RSABE/ABEL in a TSD: Science fiction

Hi Yung-jin,

» […] It sounds like that it is possible to implement all these using R or into bear together with PowerTOST.

Yes, why not?

» OK, the example is from WLN.

The example is from Potvin’s paper.

» CVintra = 18.21% (for Cmax?) with 12 subjects.

They didn’t state which PK metric the data covers.

» So it is not a HVD, isn't it?

The CV of the pooled data set (n=20) is 21.67%. So HVD is unlikely.

» Does EMA or FDA accept the two-stage BE if it is the case?

Hopefully only in such cases (conventional unscaled ABE). I have heard that some people already naïvely unreflecting submitted protocols in the EU with replicate designs for ABEL, believing that αadj 0.0294 would maintain the overall risk type I ≤0.05. Crap.

Cheers,
Helmut Schütz

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
MGR
★

India,
2014-07-02 07:08

@ Helmut
Posting: # 13199
Views: 6,746

## WinNonlin 5.2.1 setup

Hi Helmut,

Here in our company WinNonlin 5.2.1 is installed, So i have to do analysis using the same version. Done the analysis, but the problem is that, 94.12% confidence interval values are coming same when subject(sequence) is treated as fixed effect and another with random effect.

Could be please explain why this happened?

I have attached the screen shot of the settings used.

Edit: Subject line changed. [Helmut]

Regards,
MGR
Helmut
★★★

Vienna, Austria,
2014-07-02 10:40

@ MGR
Posting: # 13201
Views: 6,829

## outdated software; mixed vs. fixed effects

Hi MGR,

» Here in our company WinNonlin 5.2.1 is installed, So i have to do analysis using the same version.

I don’t know how inspectors would judge the fact that your company is using an obsolete soft­ware (which is no more supported by Pharsight since 2011). The upgrade to Phoenix was free of charge for years. Talk to your management!

» Done the analysis, but the problem is that, 94.12% confidence interval values are coming same when subject(sequence) is treated as fixed effect and another with random effect.

As expected…

» Could be please explain why this happened?

The CI is calculated from the geometric mean ratio, the residual error, and the degrees of free­dom (in a 2×2 cross-over n–2, where n is the total number of subjects). These values are iden­ti­cal in both models.1 My results for the first stage of Potvin’s Example 2, Method B in Phoenix / Win­Nonlin 6.3 – agreeing with what the authors reported:
• Fixed Sequence + Treatment + Period    Random Subject(Sequence)

Ratio:     108.76% 94.12% CI:  92.93 – 127.28% 

• Fixed Sequence + Treatment + Period + Subject(Sequence)

Ratio:     108.76% 94.12% CI:  92.93 – 127.28% 
“Philosophically” speaking treating subjects as a random effect allows prediction of the popu­la­tion’s expected response based on the study’s sample. Treating subjects as a fixed effect makes only a statement about subjects in this particular study. See also this post.

EMEA. The European Medicines Evaluation Agency.
The drug regulatory agency of the European Union.
A statistician-free zone.
Stephen Senn (2004)

However, for the EMA you have to stick to the all fixed effects model with Subject(Sequence).2 My ex­pe­ri­ences from a last year’s MRP (mixed effects model, Method C passed BE in the first stage): Accepted by the RMS Germany; no comments from the CMS Austria, Denmark, Sweden, and The Nether­lands. Nonetheless, Spain:

“Statistical analysis should be GLM. Please justify.”

Can you guess the outcome of the all fixed effects model presented in the response letter?

1. Identical both for balanced (n1=n2) and imbalanced (n1≠n2) studies. However, dif­fe­rent re­sults for in­com­plete data (missing periods). In SAS-lingo Proc GLMProc Mixed. Not an issue here (according to the EMA incomplete subjects have to be excluded in 2×2 cross-over studies).
2. For aficionados: ElMaestro does not get tired pointing out that sub­jects are un­am­bi­gu­ously coded to se­quen­ces in cross-over BE studies. Therefore, nesting Subjects into Se­quen­ce is super­flu­ous. Homework: Try what happens with the more simple Sequence + Treatment + Period + Subject.

Cheers,
Helmut Schütz

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
MGR
★

India,
2014-07-03 06:32

@ Helmut
Posting: # 13206
Views: 6,638

## outdated software; mixed vs. fixed effects

Hi HS,

Thank you very much for the explanation

Regards,
MGR
Bioequivalence and Bioavailability Forum |  Admin contact
19,567 posts in 4,150 threads, 1,341 registered users;
online 12 (0 registered, 12 guests [including 10 identified bots]).
Forum time (Europe/Vienna): 10:39 CEST

Power. That which statisticians are always calculating
but never have.    Stephen Senn

The BIOEQUIVALENCE / BIOAVAILABILITY FORUM is hosted by
Ing. Helmut Schütz